In this work we aim to extend background subtraction techniques to work in more dynamic and varied situations, which are required for surveillance and tracking algorims in natural, real world environments. The techniques is based on a hierarchy oflocal spatio-temporal models defining the background appearance at each pixel. We find it is possible to create powerful classifiers (locally) even for "backgrounds" containing complicated, non-uniform motions. Previous works considered the following models to characterize the appearance at a pixel:
These model the variations in the intensity (and can be extended to the color of the background). However, they suffer in cases of consistent background motion. To counter this, our model includes also:
Intensity (mean and standard deviation, non-parametric models)
Linear Prediction based upon time history
Optic Flow (mean and standard deviation) Multiple optical flows (adaptive mixture model) Spatio-Temporal image derivative distribution (represented as a multi-variate Gaussian, Gaussian mixture models built from EM, adaptive mixtures, and non-parametric models.)
The efficacy of these models has been demonstrated using ROC curves. More concrete measures of the power of these statistical models to distinguish foreground from background is codified as a modified relative entropy measure.
|
|
|
| Lake with wave and grass motion. sampleOutput | Traffic Intersection sampleOutput |